Article
Reference
Meta-analysis of action video game impact on perceptual, attentional, and cognitive skills
BEDIOU, Benoît, et al.
Abstract
The ubiquity of video games in today's society has led to significant interest in their impact on the brain and behavior and in the possibility of harnessing games for good. The present meta-analyses focus on one specific game genre that has been of particular interest to the scientific community—action video games, and cover the period 2000–2015. To assess the long-lasting impact of action video game play on various domains of cognition, we first consider cross-sectional studies that inform us about the cognitive profile of habitual action video game players, and document a positive average effect of about half a standard deviation (g = 0.55). We then turn to long-term intervention studies that inform us about the possibility of causally inducing changes in cognition via playing action video games, and show a smaller average effect of a third of a standard deviation (g = 0.34). Because only intervention studies using other commercially available video game genres as controls were included, this latter result highlights the fact that not all games equally impact cognition.
Moderator analyses indicated that action [...]
BEDIOU, Benoît, et al . Meta-analysis of action video game impact on perceptual, attentional, and cognitive skills. Psychological Bulletin , 2018, vol. 144, no. 1, p. 77-110
DOI : 10.1037/bul0000130
Available at:
http://archive-ouverte.unige.ch/unige:101941
Disclaimer: layout of this document may differ from the published version.
1 / 1
Meta-Analysis of Action Video Game Impact on Perceptual, Attentional, and Cognitive Skills
Benoit Bediou
Université de Genève
Deanne M. Adams and Richard E. Mayer
University of California, Santa Barbara
Elizabeth Tipton
Teachers College, New York, New York
C. Shawn Green
University of Wisconsin–Madison
Daphne Bavelier
Université de Genève
The ubiquity of video games in today’s society has led to significant interest in their impact on the brain and behavior and in the possibility of harnessing games for good. The present meta-analyses focus on one specific game genre that has been of particular interest to the scientific community—action video games, and cover the period 2000 –2015. To assess the long-lasting impact of action video game play on various domains of cognition, we first consider cross-sectional studies that inform us about the cognitive profile of habitual action video game players, and document a positive average effect of about half a standard deviation (g⫽0.55). We then turn to long-term intervention studies that inform us about the possibility of causally inducing changes in cognition via playing action video games, and show a smaller average effect of a third of a standard deviation (g⫽ 0.34). Because only intervention studies using other commercially available video game genres as controls were included, this latter result highlights the fact that not all games equally impact cognition. Moderator analyses indicated that action video game play robustly enhances the domains of top-down attention and spatial cognition, with encouraging signs for perception. Publication bias remains, however, a threat with average effects in the published literature estimated to be 30% larger than in the full literature. As a result, we encourage the field to conduct larger cohort studies and more intervention studies, especially those with more than 30 hours of training.
Public Significance Statement
Understanding the effects of action video game play is essential given that (a) a large number of individuals regularly spend many hours on these types of games, and (b) proponents are offering suites of video games that are claimed to change behavior or enhance cognition. The 2 meta-analyses in this paper present the current status of this field, concluding that playing action video games has some positive effects on improving cognitive skills. This review also identifies some limitations of current research.
Keywords:action video games, attention, cognition, meta-analysis, perception Supplemental materials:http://dx.doi.org/10.1037/bul0000130.supp
This article was published Online First November 27, 2017.
Benoit Bediou, Faculté de Psychologie et Sciences de l’Education (FPSE), section de Psychologie, Université de Genève; Deanne M.
Adams and Richard E. Mayer, Department of Psychological and Brain Sciences, University of California, Santa Barbara; Elizabeth Tipton, Department of Human Development, Teachers College, New York, New York; C. Shawn Green, Department of Psychology, University of Wisconsin–Madison; Daphne Bavelier, Faculté de Psychologie et Sciences de l’Education (FPSE), section de Psychologie, Université de Genève.
We thank Brett Ouimette for his assistance with gathering and reviewing articles. We are also grateful to our librarian Dominique Vallee for her help with the literature search. We thank all members of the Bavelier lab for their help with search in non-English languages, as well as Ekatarina Plys, Anna-Flavia Di Natale, and Sabine Öhlschläger, who helped with screening of inclusion criteria and coding of study variables, and Nuhamin Petros for her assistance with formatting of the manuscript references. We are especially thankful to Zhipeng
Hou for his invaluable help in working out the confidence intervals for multiple moderator models and the moving constant technique approach in R. This project was supported by Grants 100014_159506 and 100014_140676 from the Swiss National Science Foundation and N00014-14-1-0512 from the Office of Naval Research (ONR) to Daphne Bavelier as well as Grant N00014-11-1-0225 from the ONR to Richard Mayer and Grant N00014-17-1-2049 from the ONR to Shawn Green.
Data and code are kept available online on request (https://osf.io/3gd8v/).
Daphne Bavelier declares she is a member of the scientific advisory board of Akili Interactive, Boston.
Correspondence concerning this article should be addressed to Benoit Bediou or Daphne Bavelier, Faculté de Psychologie et Sciences de l’Education (FPSE), section de Psychologie, Université de Genève, Bou- levard pont d’Arve, 40, CH-1205 Genève. E-mail: benoit.bediou@
unige.chordaphne.bavelier@unige.ch ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
77
Over the past 20 years, there has been significant scientific interest in examining the potential behavioral consequences of playing video games. This research is still in its relative infancy, but one repeatedly observed finding, from social psychology to clinical psychology to educational psychology, to the focus of this meta-analysis, cognitive psychology, is that not all video games have the same impact. Indeed, given the enormous range of com- pletely different experiences that fall under the label of video games, attempting to identify how playing video games affects behavior is analogous to attempting to identify how eating food impacts physiology. The current study therefore focuses on one specific game genre, termedaction video games, for which there are now sufficient data to examine impact via meta-analytic tech- niques, covering the period from year 2000 until 2015. Here we thus ask, “Do people learn anything useful from playing action video games?”
The rationale for this study is both practical and theoretical, placing it in the category of whatStokes (1997)callsuse-inspired basic research. Specifically, whether people learn anything useful from playing action video games is an important practical question in light of the large number of individuals who play action video games for extended periods of time. Indeed, according to recent reports, more than 1.2 billion individuals world-wide (including more than 150 million Americans) are video gamers, with action games consistently ranking at or near the top for most popular game type (Spil Games, 2013;The Entertainment Software Asso- ciation, 2015). Perhaps not surprisingly then, as video gaming has surged in popularity, so too has popular interest in the potential practical ramifications of such gaming on our everyday lives (Bavelier & Green, 2016). Examining the cognitive profile of habitual action video game players, who have spent hundreds of hours playing this particular type of games, is an important first step for understanding the long-lasting cognitive changes associ- ated with action video game play. This is not the only practical concern related to the impact of action video games. For instance, many research groups have started to use such off-the-shelf com- mercial action video games in translational applications (Mayer, 2014, 2016). Thus, confidence in the overall accuracy of that foundational science is critical, as has been recently highlighted in the context of brain training games (Simons et al., 2016).
It is also an important theoretical question in light of the pro- posal that the general skills learned from playing in a game context can transfer to nongame contexts that require the same underlying skills (Green & Bavelier, 2012; Mayer, 2014; Sims & Mayer, 2002). Such growing interest into the gamification of various interventions is exemplified by the recent surge of new genres of games, such as therapeutic or educational serious games, games for impact, crowdsourcing games, or ‘so-called’ brain-training games. Here, we used a meta-analytic approach to paint a more coherent and global picture of the effects of playing action video games.
Action Video Games
As noted above, there are many types of video games that can differ dramatically from one another. Video games within the action genre all share a set of qualitative features, such as: (a) a fast pace (in terms of the speed of moving objects, the presence of many highly transient events, and the need to make motor re-
sponses under severe time constraints); (b) a high degree of per- ceptual and motor load, but also working memory, planning and goal setting (e.g., many items to keep track of simultaneously, many possible goal states that need to be constantly reevaluated, many motor plans that need to be executed rapidly); (c) an em- phasis on constantly switching between a highly focused state of attention (e.g., toward aimed targets) and a more distributed state of attention (e.g., to monitor the whole field of view); and (d) a high degree of clutter and distraction (i.e., items of interest are distributed among many nontarget items). The main subtypes of games generally considered to be action video games include first-person shooter games, wherein the player views the world through the eyes of his or her avatar (e.g., theHalo, Call of Duty, and Medal of Honor series of games) and third-person shooter games, wherein the player sees the back of his or her avatar (e.g., theGears of WarandGrand Theft Auto series of games). Thus, games such as Rise of Nations, Pac-Man, or Space Fortress, although occasionally also given the label of action games in various parts of the literature, do not qualify as action games for the purposes of this meta-analysis because they lack the game mechanics described above.
It should be noted that although the action game features listed above were reasonably unique to action video games circa the year 2000, today some, albeit not all, of these critical characteristics can now be found in other genres such as multiplayer online battle arena games, real-time strategy games, or role-playing games (Dale & Green, 2017). In the present work, however, we use a narrow definition of action video game, whereby genres such as real-time strategy, role-playing, and fighting video games do not qualify, as this is the definition that has been employed in the literature to date.
Previous Syntheses of Gaming and/or Action Gaming Research
The use of the meta-analytic approach, combined with a focus on action video games and their impact on cognition, departs from other literature reviews that provide qualitative descriptions of research studies rather than measures of effect size (Connolly, Boyle, MacArthur, Hainey, & Boyle, 2012;Hays, 2005;Honey &
Hilton, 2011;Randel, Morris, Wetzel, & Whitehill, 1992;Tobias, Fletcher, & Bediou, 2015; Tobias, Fletcher, Bediou, Wind, &
Chen, 2014;Tobias, Fletcher, Dai, & Wind, 2011;Young et al., 2012) or that focus on academic learning outcomes rather than on skills (D. B. Clark, Tanner-Smith, & Killingsworth, 2014; Sitz- mann, 2011;Vogel et al., 2006). A few reviews focusing more squarely on cognition and carried out at the level of all video games have certainly provided interesting pointers (Adams &
Mayer, 2014; Connolly et al., 2012; Powers, Brooks, Aldrich, Palladino, & Alfieri, 2013;Simons et al., 2016;Tobias et al., 2011;
Young et al., 2012). Yet, it is important to consider that the label ofvideo gamesapplies equally well to experiences that may differ dramatically from one another. Experiences that fit the superordi- nate category label ofvideo gamesmay involve incredibly simple graphics or alternatively may involve highly realistic and lifelike environments, solitary single-player activities or rich social struc- tures, minutes-long play or years-on-end play, mostly reactive activities or substantial long-term planning, distinctly pro-social behavior or rather antisocial behavior, elaborate game mechanics ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
or nothing more than a set of rules, pure entertainment or primarily translational, and so forth. In short, the extreme flexibility of the media allows for all kinds of experiences. Given that the behav- ioral outcomes of an experience depend strongly on the nature of the experience (D. B. Clark et al., 2014), a superordinate category label such as video games is likely to have limited predictive power. Indeed, it seems highly unlikely that a slow-paced game with essentially no perceptual, attentional, or cognitive demands (e.g.,Farmville) would produce the same outcomes as a fast-paced game that places heavy demands on the perceptual, attentional, and cognitive systems (e.g.,Call of Duty) – despite both clearly being video games.
To date, only three published meta-analyses have examined the effect of playing specific video game genres on cognition. The first two come from the group of Powers and colleagues (Powers &
Brooks, 2014;Powers et al., 2013). In their first meta-analysis, Powers et al. (2013)included all video games and coded for game genre as a moderator (using 5 categories: action/violent, mimetic, nonaction, puzzle, nonspecific). This moderator was found to have a significant effect, indicating differences between game genres, although the exact source of that effect was difficult to pinpoint (see their Table 5 on page 1067). Interestingly, the effect of action video games was significant in both quasi-experiments (i.e., cross- sectional studies comparing selected groups of habitual players of action video games with nonvideo game players, matched for as many factors as possible other than their video gaming habits) and in true or intervention studies examining whether video game training produced changes in performance.
In their subsequent report (Powers & Brooks, 2014), the authors focused squarely on intervention studies alone. This reanalysis confirmed an effect of First Person Shooter games (more in line with our definition of action video games), and demonstrated that training with these games produced a significant overall improve- ment in cognition (d⫽0.23, 95% CI [0.07, 0.39],p⫽.005). The authors also further examined a number of subdomains of cogni- tion, with significant effects being found for perceptual skills (k⫽ 35,d⫽0.45, 95% CI [0.17, 0.72],p⫽.001), and spatial imagery (k ⫽ 11,d⫽ 0.17, 95% CI [0.01, 0.34], p⫽ .04), but not for executive functions (k⫽10,d⫽ ⫺0.17, 95% CI [⫺0.47, 0.14], p⫽.28) or for motor skills (k⫽1,d⫽0.07, 95% CI [⫺0.31, 0.45],p⫽.72).
A third and more recent meta-analysis byWang et al. (2017) arrived at the same conclusion about the positive effects of action video game training, despite using a quite different definition of action video games. In particular, Wang et al. (2017) included games such as Pac-Man that would clearly not be considered action games in the present meta-analysis or in the meta-analyses by Powers and colleagues. As different games may have different impact on cognition, the present meta-analyses stick to a more coherent definition of action video games, as described above.
The present study thus builds on these previous meta-analyses, while adding a number of novel aspects. The first notable differ- ence lies in the strict definition of action video games used in the present analyses, which imposed different constrains on our liter- ature search and selection criteria. For example, our study covers the period from year 2000 –2015, whereas Powers and colleagues (Powers & Brooks, 2014;Powers et al., 2013) covered from 1980 to 2012.
The inclusion of three extra years at the end of the range (i.e., 2013–2015) is of clear importance, as the rapid growth of publi- cations on video games over the past few years means that 48% of the studies we included in the current analyses (2013–2015) were not included inPowers et al. (2013,2014). However, the differ- ence in start date is perhaps even more important. In particular, by considering only work starting with the year 2000, we ensure that we are categorizing under the action video game genre relatively homogenous game experiences. Indeed, games that are recogniz- able in terms of modern action mechanics only came into being in the late-1990s (this includes acclaimed titles such asDoom(1993), Goldeneye (1997), Half-Life (1998), and Counterstrike (1999).
Although games today have obvious graphical and computational advantages as compared with the games of the late 1990s and early 2000s, the core components are nonetheless strongly shared across this range (i.e., most of the base mechanics as related to movement and aiming are essentially identical). This is not true though of games from the early 1980s because most action-like games from that era may include games such as Centipede or Super Mario Bros., which bear little resemblance in terms of action mechanics and content to modern action games. Thus, lumping games to- gether from as early as the 1980s will also tend to confound efforts to examine the impact of what are now considered key action game characteristics.
We have also taken special care to not conflate action and violent video games. A relatively common misbelief about the world of gaming is that these terms are interchangeable; yet they are not. Simply put, not all action games contain violence and not all games that contain violence are action games. It is undeniably the case that most first-person and third-person shooter games are violent. Yet, violence is not essential to the action mechanics we describe above. As such, it is perfectly possible for a game to utilize action mechanics and dynamics in the absence of any violent content (e.g., as the case in cooperative paintball games such asSplatoon, or child friendly shooter games such as Ray- man’s Raving Rabbids). It is also equally possible to have violent content in games without any action characteristics (e.g., many turn-based role playing games, such asFinal Fantasy VII, have a great deal of violence in the absence of any action components).
This latter point is particularly critical with respect to the creation of a combined action/violent category, as the presence of violent, but not action games in an analysis can potentially confound attempts to isolate the effect of action games.
Our meta-analyses further departs from previous meta-analyses on the impact of action video games by restricting intervention studies to only those that used other commercially available game genres as controls. This point is critical if we are to understand the specific features of action games that impact cognition. It also sets a high bar for observing an effect, as intervention studies with well-matched active controls are known to result in smaller effect sizes (Uttal et al., 2013). For example, our meta-analytic approach includes studies comparing the impact of action video games to that of Tetris on spatial cognition, when Tetrishas been docu- mented to improve spatial cognition (Uttal et al., 2013).
Furthermore, another unique aspect of our meta-analytic approach is to focus on the long-lasting impact of action video game play on cognition. In particular, intervention studies were restricted only to those studies that tested impact at least 24 hours after training com- pletion, thus ruling out a number of potentially fleeting confounds ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
(e.g., improvements attributable only to arousal). As long-lasting plastic changes are notoriously hard to induce, our meta-analysis is quite representative of the field by also requiring at least 8 hours of training distributed over 8 days.
Finally, the present meta-analysis makes use of recent meta- analysis methods (robust variance estimation, hierarchical model, multiple moderator analysis) to examine a host of previously unaddressed methodological issues both with respect to the action gaming literature itself, such as the impact of moderators related to motivation or expectation biases, as well as with respect to issues unique to meta-analyses, such as publication bias and small study effects (see below for additional description).
Key Issues to Be Considered
Which Cognitive Domains Are/Are Not Impacted by Action Gaming?
The extent to which a given experience alters behavior should depend strongly on the match between the experience and the neural process underlying the measured behavior as well as the extent to which the processes are themselves plastic. The mechan- ics and dynamics inherent in action games do not place equivalent load on all cognitive domains (e.g., strong load is placed on processes related to top-down attention, perception, and multitask- ing, but there is little to no verbal cognition at play;Spence &
Feng, 2010). Therefore, it would be surprising if all cognitive domains were equally altered— or altered at all— by action gam- ing. We use the broad termcognitive domainto refer to all aspects of cognition, including perceptual skills, attentional skills, and cognitive skills.
Understanding the cognitive domains that are and are not modified by action gaming is not only important for our theoretical understand- ing of how action game play promotes behavioral changes, but is also particularly crucial for those researchers attempting to utilize action games for practical ends. For example, some reports of action video game play improving mental rotation (Feng, Spence, & Pratt, 2007), have led to the proposal that these games may be useful for education in the sciences, technology, engineering, and mathematics disciplines (Uttal et al., 2013). Similarly, reports of faster and more accurate visuomotor control after action game play has led to studies probing their usefulness to train laparoscopic surgeons to perform surgeries faster without making more errors (Schlickum, Hedman, Enochsson, Kjellin, & Fellander-Tsai, 2009). Reports of enhanced perceptual (primarily visual) processing after action gaming has led to studies assessing their utility in the rehabilitation of visual disorders such as amblyopia (J. Li et al., 2015;R. W. Li, Ngo, Nguyen, & Levi, 2011;
Vedamurthy, Nahum, Bavelier, & Levi, 2015). Finally, reports of enhanced visual attention after action game play have led researchers to use these games to train Italian dyslexic children, for whom attention appears to be one of the major bottlenecks that constrains the fluency of reading (Franceschini et al., 2013).
Given this strong movement toward translational work, it would thus be reassuring to validate that these games have an impact on cognition, and identify which skills may be more reliably affected.
To this end, we categorized measures of cognition into the follow- ing eight cognitive domains, guided by available data: (a) percep- tion (e.g., contrast sensitivity, lateral masking), (b) bottom-up
attention (e.g., pop-out search, exogenous cueing), (c) top-down attention (e.g., complex search, flanker tasks, multiple object tracking), (d) spatial cognition (e.g., mental rotation, spatial work- ing memory tasks), (e) task-switching/multitasking (e.g., dual-task or task-switch paradigms,), (f) inhibition (e.g., go-nogo, stop- signal tasks, proactive interference), (g) problem solving (e.g., Tower of Hanoi, Tower of London, Raven’s matrices), and (h) verbal cognition (e.g., verbal working memory, reading).
Type of Study Design: Cross-Sectional and Intervention Studies
Studies examining the long-lasting effects of action gaming on cognition have typically taken one of two forms. The first type are cross-sectional designs. Here, the performance of self-selected individuals who naturally play a large amount of action video games (often labeled to asaction video game playersor AVGPs) is contrasted with the performance of individuals who specifically do not play those kinds of fast-paced action-packed video games and rarely, if at all, play other nonaction types of games (often referred to asnonvideo game playersor NVGPs). Although these groups differ in their video game play, they are matched along as many potentially confounding dimensions as possible, including factors such as age-range, gender, and years of education. The critical measure in these studies is thus related to a difference in performance between these two extreme self-selected groups (i.e., whether AVGPs show better performance than NVGPs).
Although self-selection bias is always a concern for such cross- sectional studies, they nonetheless serve a dual role in the charac- terization of the long-lasting effects of action video game play on cognition. First, they document the cognitive profile of a growing segment of the population (AVGPs), an important societal ques- tion. Second, they provide a useful pointer as to where it may be worth investing in a training study. Indeed, documenting different cognitive skill between AVGPs and NVGPs clearly calls for an intervention study. In contrast, if hundreds of hours of action game play over months to years do not result in a group difference in the cognitive skill tested, the expectation that durable changes of that cognitive skill will be observed after only tens of hours of video game play is lessened. Accordingly, 65% of the intervention studies identified in the literature first established a cross-sectional effect of action video game play.
The second broad type of study involves intervention studies (also called true experiments). Here, individuals who do not, as part of their normal life, tend to play much video games are specifically trained on either an action video game or a control video game with performance on the skills of interest being mea- sured both before and after training. The critical measure in these studies is thus related to a difference of differences—specifically assessing whether the action trained group showed greater im- provements from pretest to posttest than the control trained groups.
Among intervention studies, we focus exclusively on those which contrasted training on an action video game with training on a control video game. As is standard in the field, the experimental action and the control, nonaction games were required to be commercially available games, ensuring that both arms of the study were engaged in a quality, entertaining and challenging video game experience. Studies using, either no control, passive controls or repeated practice on the task of interest (e.g., group ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
receiving Useful Field of View [UFOV] training inBelchior et al., 2013), are not included.
Focusing exclusively on intervention studies with an entertain- ment quality video game control group makes the present meta- analysis quite unique in comparison to those of others (Powers et al., 2013;Powers & Brooks, 2014;Wang et al., 2017). As noted previously, it also sets a high bar for the magnitude of effects that must be produced by action video game training. Indeed, not only is it the case that recent work indicates that some game genres that have, in some studies, been used as a control may also enhance cognition (Blumen, Gopher, Steinerman, & Stern, 2010; Glass, Maddox, & Love, 2013;Powers & Brooks, 2014;Powers et al., 2013;Wang et al., 2017), it is also the case that, more generally, effect sizes of interventions with well-matched active control are typically smaller than those obtained when including all interven- tion studies (Uttal et al., 2013). Thus, the present meta-analysis of intervention studies is a departure from most, if not all intervention meta-analytic work published thus far as the control groups con- sidered here go well beyond a simple placebo group (e.g., being given an inert pill) to control for possible expectation effects.
Instead, many of the video games used as controls provide rich immersive experiences that are likely to have an impact of their own, whether on cognition or on other aspects of behavior.
Importantly, active control intervention studies fall under the label of what Mayer (2011, 2014) has called cognitive conse- quences research,which is essential in defining the impact of the video game genre studied on various cognitive outcomes, as our aim is here (see alsoBoot & Simons, 2012;Green, Strobach, &
Schubert, 2014;Jacoby & Ahissar, 2013;Schellenberg & Weiss, 2013for best practices in behavioral intervention studies). Unlike in cross-sectional studies, all participants in an intervention study are recruited using the same method, and participants are then randomly assigned to their treatment group. Therefore, recruitment method is not a relevant moderator for these types of studies.
Conversely, training duration only applies to intervention studies.
These constraints determine the moderators considered below dur- ing the analysis.
A number of factors are likely to moderate the effects of action games, some of which relate to participant characteristics (e.g., age), whereas others relate to methodological aspects (e.g., recruit- ment, type of measure, or duration of training). One of our aims is to take advantage of the meta-analytic approach to examine how different factors may alter the impact of action video game play.
Participant Age
Most available studies examining the impact of action video games on cognitive function have focused on college-aged indi- viduals. A few studies, though, have examined the effect of action video game play in normal children (under age 18) or in older adults. Although the current literature is not ideal with respect to exploring effects from a life span perspective, in that there are no cross-sectional studies including older adults and no intervention studies including children, we can nonetheless look at possible age effects within each of these types of studies.
In general, plastic changes are typically greatest in children and then decrease in magnitude with aging. Recent work though high- lights the potential for brain plasticity throughout the life span, even into old age (Mahncke, Bronstone, & Merzenich, 2006).
Provided that the stimulation is of appropriate difficulty and spe- cifically targeted toward the to-be-enhanced skills, it seems computer- or game-based training induces small sized benefits in older adults, especially in verbal memory, speed of processing, and verbal/spatial working memory with the impact on attention and executive functions being less reliable (Ball et al., 2002; Karr, Areshenkoff, Rast, & Garcia-Barrera, 2014;Lampit, Hallock, &
Valenzuela, 2014;Toril, Reales, & Ballesteros, 2014;Wang et al., 2017). In contrast to the experiences above, which were carefully titrated to the abilities of older participants, action video games are designed to be challenging for young individuals, and perhaps more to the point, young individuals who are already well versed in the demands of action games. Not surprisingly, such games are typically too challenging for older adults. As a result, very few studies have used action video games as defined above as an intervention tool in older adults (i.e., 2 records; 12 effect sizes).
Given the importance of providing users with a training regimen within their proximal zone of development to induce learning, one can expect rather different outcomes of training with an action video game in young and older adults. The present analysis is poised to shed some light on this issue.
Type of Dependent Measure
The most common dependent measures in cognitive psychology are accuracy and reaction time (RT). Many tasks involve collecting only one or the other. Even in those tasks that nominally measure both accuracy and RT, it is typically the case that one is the primary measure of interest, with the other being used mainly to rule out potential confounds such as speed–accuracy trade-offs. Of interest here is whether RT or accuracy is differentially affected by action video game play. Specifically, by examining whether the type of dependent variable used in studies moderates the size of the action video game play effect, we can test the popular belief that action video game play involves a certain degree of “trigger happy” behavior, whereby speed is valued over accuracy. If so, we would expect to see larger effect sizes in speed as compared with accuracy measures. If, however, the effects of action games are truly at the level of the core processes themselves, then we would expect to see effects on both accuracy and RT.
Main Versus Difference Effects
There have been a number of recently published criticisms suggesting that the effects attributed to action video game experi- ence in the literature could instead be ascribed to participant expectation effects (Boot, Blakely, & Simons, 2011;Boot, Simons, Stothart, & Stutts, 2013;Kristjansson, 2013, but seeBisoglio et al., 2014;Green et al., 2014). A first foray in assessing thisexpectation bias hypothesis has led us to separately code studies where the effect of interest is a main effect (e.g., overall change in RT or accuracy) versus studies where the effect of interest is a difference between conditions (e.g., disproportionately faster RTs or higher accuracy in some conditions vs. others). Main effect hypotheses such as “action game players should respond faster” are more likely for a naïve participant to intuit, as well as potentially easier to match behavior to, than a difference effect hypotheses such as
“action game players should respond disproportionately faster on switch trials than on non-switch trials.” Thus, if action video game ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
effects result from participants trying to match their behavior to what is expected of their assigned group, we would predict larger group differences in studies where the effect of interest was a main effect and smaller group differences in cases where the effect of interest was a difference term. The strength of action video game play on difference effects will thus provide an indirect assessment of the validity of the expectation bias hypothesis.
Recruitment Method in Cross-Sectional Studies
Although the analysis of main versus difference effects is par- tially relevant to this point, a more direct assessment of the as-yet-untested expectation hypothesis comes from the analysis of recruitment methods. Specifically, many studies comparing action video game players with nonaction players in the literature have employed overt recruitment methods (e.g., posters asking about gaming habits). Because the participants in these studies know that they have been selected based upon their gaming habits, expecta- tion effects are possible. Other studies though have employed purely covert recruitment methods (i.e., participants do not know they are being selected based on their gaming habits). Expectation effects are unlikely, because the participants in these studies do not know that gaming habits are of interest. If knowledge of one’s video game status drives the entirety of the reported effects on action video game play (which we label here as the most extreme form of the expectation hypothesis), then no effect of action gaming status should be observed in studies that have employed covert methods.
Training Duration in Intervention Studies
Intervention studies vary in terms of training duration. Because time-on-task is a significant predictor of learning, studies using short training durations may report smaller effects than studies with longer training duration. Here we used metaregression meth- ods to test whether training duration is linearly related to the strength of action video game effects.
Laboratory
Most of the earliest work around the topic of action games was performed by the Bavelier laboratory. Although work from this laboratory now makes up a considerably smaller total percentage of the literature, it is nevertheless the case that whenever a single group contributes a large number of data points to a meta-analysis, it is important to consider how well effects generalize across laboratories.
The Current Meta-Analyses
In sum, the present meta-analyses focus on the impact of action video games on behavior considering a range of cognitive domains and does so separately for cross-sectional and intervention designs.
We also address timely issues in the field such as the relative impact of game play on measures of RTs versus accuracy, poten- tial confounding factors such as those related to the expectation bias hypothesis, as well as the impact of training duration in intervention studies.
Our meta-analyses thus not only extend prior meta-analyses (e.g., Powers & Brooks, 2014;Powers et al., 2013;Toril et al.,
2014;Wang et al., 2017), but they also depart from previous work on methodological and theoretical grounds. First, our meta- analyses focus on the action video game genre defined based on specific mechanics likely to have differential impact across cog- nitive domains. This is important, as both the two meta-analytic studies ofPowers and colleagues (2013,2014) and our own studies (Cohen, Green, & Bavelier, 2007) have indicated that not all video games have the same impact on different aspects of cognition.
Second, we perform separate meta-analyses of cross-sectional and intervention studies. By including only intervention studies with active controls that made use of commercially available video games, the current analysis of action video game impact is unique in addressing a number of confounding variables related to possi- ble differences in novelty, engagement, motivation, and fun, as both experimental and control groups are faced with a commercial grade experience. Third, we provide a more fine-grained analysis of impact across different domains of cognition, which diverges from the classification chosen by Powers and colleagues, and includes additional domains such as top-down attention and verbal cognition. Fourth, we examine the impact of a number of moder- ators, in particular that of possible expectation biases which have been the focus of a number of recent critiques of this work, and were not addressed in previous meta-analyses. Finally, we take full advantage of the hierarchical structure of our data set and invested specific effort to explore methodological issues related to small study effects and moderator effects.
Method Study Selection
The literature search covered the period between January 2000 and November 2015. We queried the databases PsycINFO, PsyINDEX, ERIC, FRANCIS, MEDLINE, SCOPUS, Web of Science and ScienceDirect, using terms combined in the following Boolean expression (“video game” OR “computer game”) AND (“attention” OR “attentional” OR “attend” OR “cognitive” OR
“cognition” OR “perception” OR “perceptual”), or else repeating the search with different combinations when the Boolean search was not permitted. Databases were either interrogated using their web interface, or using multidatabase search interfaces such as Ovid, EBSCO and PROQUEST. Results from the database PsycINFO can be viewed via the link provided in the supplemen- tary material.
The validity of a systematic review or meta-analysis is highly dependent on the underlying data, and more specifically the ability to reduce the potential sources of publication bias by including data from unpublished sources (Rothstein, Sutton, & Borenstein, 2005). One significant concern in this endeavor is the search for what is generally labeled asgray literature(Borenstein, Hedges, Higgins, & Rothstein, 2009a;Mahood, 2006;Rothstein & Hopewell, 2009), which com- prises work that is not published in research journals covered by these databases or not published at all. Although the Internet provides increasing access to information about published and unpublished studies, the search for, and inclusion of, gray literature remains a challenge and is still unanimously recognized to be particularly dif- ficult (Hopewell, Clarke, & Mallett, 2006).
Thus, in an effort to be as exhaustive as possible, we took a number of steps to identify potential sources of unpublished stud- ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
ies. First, we identified possible sources of gray literature related to the impact of (action) video games on perception, attention, or cognition, by following the recommendations found in various books, articles and websites dedicated to publication bias and gray literature, such as the Cochrane Handbook and the Library of the University of Western Australia. Given these recommendations we: (a) queried several databases that specialize in gray literature (e.g., PsycEXTRA, ScholarOne, opengrey, base-search); (b) searched the abstracts of the annual conferences of the Society for Neuroscience, the Vision Science Society, the Cognitive Neuro- science Society, and the annual convention of the American Psy- chological Association (APA); (c) conducted additional searches in Google Scholar, and in the database Dissertations Abstracts International; (d) posted to a number of relevant listservs includ- ing those for the Vision Sciences Society (VSS) and the Color &
Vision Network (CVNet); and (e) directly contacted 67 authors who are known to have worked on the topic and asked whether they had or were aware of unpublished data examining the effects of video games on attention, perception or cognition.
We further extended the literature search to include other lan- guages: Chinese, French, German, Italian, Portuguese, Romanian, Russian, and Spanish. The literature search in non-English lan- guages was performed using the same databases (but with the keywords translated into the corresponding language), as well as additional tools specific to each language, such as tesionline (Ital- ian), wanfang (Chinese), plural (Romanian), and the National Library of Romania to cite a few. Whenever possible, we confined the search to examine only the title, abstract, subject, and keyword fields. Only if a paper’s potential for inclusion was unclear was the full text consulted. We also limited our search to relevant disci- plines (e.g., psychology, neuroscience, computer science, educa- tion, sociology), and only included written material (e.g., books, chapters, articles, reviews), and thus discarded other types of documents (e.g., video, audio, biography), as well as irrelevant subjects or topics (e.g., academic guidance counseling, music, mental health, robotics, nutrition).
Throughout the search process, we paid particular attention to literature reviews or comments (Achtman, Green, & Bavelier, 2008;Bavelier et al., 2011;Bavelier, Green, Pouget, & Schrater, 2012;Bisoglio et al., 2014;Boot et al., 2011;Boot & Simons, 2012;Boot, Simons, et al., 2013;Granic, Lobel, & Engels, 2014;
Green & Bavelier, 2012;Kristjansson, 2013;Latham, Patston, &
Tippett, 2013;Oei & Patterson, 2014;Spence & Feng, 2010), even when the focus was not directly relevant to the present meta- analysis (Connolly et al., 2012; Sitzmann, 2011; Tobias et al., 2011;Vogel et al., 2006;Wouters, van Nimwegen, van Oosten- dorp, & van der Spek, 2013; Young et al., 2012), to ensure all potentially relevant references were examined for inclusion. We further cross-checked with recently published meta-analyses (Powers & Brooks, 2014;Powers et al., 2013;Wang et al., 2017), one of which examined a much larger range of video game related research. Finally, the reference lists of all the documents that passed the inclusion criteria were also consulted.
Our search, combining the results of all databases and languages as well as all keywords combinations described above, yielded a total of 958,147 hits, including 48 records in Chinese, 2,360 in French, 59,360 in German, 11,575 in Italian, 120,012 in Portuguese, 26 in Romanian, and 29,720 in Spanish. As expected, there was substantial redundancy across databases. After removal of duplicates, the remaining 676,102
references were screened by 4 raters (two authors of this article and two graduate assistants), who only read the titles of the articles (as well as the tables of contents in case of books and theses) and were trained to exclude studies that fell outside the scope of the present meta-analysis. Studies were excluded if they were only theoretical, did not involve a group of video game players or video game training, if they included only patients, or if they did not involve a measure of perception, attention or cognition. Most of the studies excluded at Step 1 dealt with topics such as the relationship(s) between video game habits (e.g., frequency/types of games played) and sociodemo- graphic information, measures of psychopathology, or of social/per- sonality factors (in particular measures related to aggression). In addition, we noted a substantial body of literature focused on the development and potential impact of serious video games for educa- tional purposes, as well as a growing literature on the use of video games as therapeutic tools designed for targeted clinical populations.
Lastly, a large body of work examining which game characteristics can increase the motivation to play or induce a flow experience was also excluded.
The 5,770 documents that remained after Step 1 were then reduced to 630 documents in Step 2 after a reading of the abstracts. Finally, after careful reading of the methods sections of these sources in Step 3, we further excluded 549 references that did not meet the inclusion criteria detailed in the next section below (which also describes representative examples of studies excluded at this stage and the reason(s) they were excluded). The final dataset thus contained 82 studies, including 65 published studies (all in English) and 17 unpub- lished studies. The 17 unpublished studies consisted of 6 doctoral theses (five of which were in English and one in Spanish), 3 personal communications, and 8 posters from various conferences. Several of these records contained both cross-sectional and intervention studies.
Of the 73 records with cross-sectional designs, 15 were unpublished (20%), whereas of the 23 records for intervention studies only 2 were unpublished (9%). It is worth noting that a substantial proportion of the unpublished work identified in our initial search contained data that were later published (and is utilized in the published form here), a fact that likely lowered the prevalence of unpublished work included in this meta-analysis.
Throughout all steps of the process (from literature search to study selection, effect size computation and coding of moderators and other study descriptors) authors were contacted to obtain missing informa- tion. In total, 67 authors were contacted for various reasons (e.g., missing data, clarification of method, etc.) and all of these authors were, at the same time, asked about potential unpublished work.
Authors from one paper were sometimes contacted together, resulting in 51 e-mails being sent. We obtained 47 responses of 51 requests (92% response rate), and authors sent the requested data or informa- tion in 40 cases (85% success), which led to inclusion in 30 cases (78% inclusion). Studies were excluded if the information provided by the authors indicated that the study did not fit our inclusion or exclusion criteria (10 studies) or if the authors were not able to provide the requested data (7 studies), either because it was not collected or because it was no longer available.
Selection Criteria
The 2nd and 3rd steps of our study selection consisted of reviewing the abstracts and/or methods sections of the 5,770 eligible sources respectively, to verify if the sources satisfied the ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
inclusion criteria. In short (see below for additional detail), in Step 2, studies were rapidly examined to determine whether they could potentially be relevant to our meta-analysis. Then, in Step 3, the 693 remaining studies were examined more carefully to ensure that the experimental design met all our inclusion/exclusion criteria.
Three authors and two research assistants were involved in this process (with the research assistants primarily playing a role in Step 2). In all, each study that was eventually included was thus processed by at least two persons. While most studies were easily categorized as meeting or failing to meet the inclusion require- ments, those few studies that were less clear were put aside and discussed during group meetings of at least three of the authors until unanimous consensus was reached.
Step 2 - Quick scan: A study passed through Step 2 if it included a comparison between a group that engaged in action video play and a control group, on a measure of attentional, perceptual, or cognitive skill, or if the performance in one of these domains was measured before and after a video game training period with an action video game trained group being contrasted with a nonaction video game trained group. Studies that measured other cognitive skills, such as risk taking or delay discounting (e.g.,Bailey, West,
& Kuffel, 2013), and studies contrasting different types of expe- rienced video game players, such as action and role playing in Krishnan, Kang, Sperling, and Srinivasan (2013), were thus ex- cluded at this stage. Although these criteria were often easily verifiable by examining the abstracts, we generally also screened the methods sections to check for whether relevant background measures such as verbal IQ could nevertheless be extracted. For example, in the study byBailey and colleagues (2013)introduced earlier, the authors mentioned that the participants also completed the useful-field-of-view and stop-signal tasks. Thus, while the risk taking data did not fit any of the cognitive domains investigated here, the UFOV and stop-signal data could still be included; these data were not available in the paper and were obtained from the thesis of Benoit Bediou (Bailey, 2012).
Step 3 - Thorough screening: After careful reading of the full text, a study (or where applicable specific individual experiments within the manuscript under consideration) was included if it satisfied all the five inclusion/exclusion criteria detailed below.
Illustrative examples of excluded studies and the reason(s) for exclusion can be found in supplementary Table S1.
1. Effect size measure. The study measured performance in one (or more) of the 8 domains of cognition introduced earlier, in healthy participants, even if this measure was not the primary dependent measure in the study. Studies were included only if we could extract measures of mean and standard deviation or recover other information permitting the calculation of effect size.
2. Game genre and hours of practice criteria for cross- sectional studies. Our focus on action video games means that predominantly only first- and third- person shooter games were included. This definition excluded not only games that clearly belong to other genres, such as real-time strategy games (e.g., Starcraft, Rise of Nations), puzzle games (e.g.,Tetris, Portal), or role-playing games (e.g.,World of Warcraft), but also games that may have been classified as action games by some authors, such as fighting games (Tanaka et al., 2013), arcade games like Pac-Man, or the cognitive research interface Space Fortress (Wang et al., 2017).
Importantly, this definition also excluded all studies conducted before the year 2000, because at that time the existing power and graphics
limitations did not allow for the fast and complex dynamics charac- teristics of action games described earlier.
Studies comparing habitual action video game players to nonaction video game players were included if the participants in the AVGP group played at least 3 hours per week of action video games and had done so for the last 6 months,andparticipants in the NVGP group spent less than 1 hour per week playing specifically action video games, or played fewer than 3 hours per week of video games in general, across all genres. These criteria were used as they capture most of the studies in the field. Some laboratories, however, have systematically used more stringent criteria, like the Bavelier labora- tory which has always required at least 5 hours per week of action video game for the AVGP group and no more than 1 hour per week of play in other genres for NVGP.
Studies were thus excluded either because the AVGP group did not play a minimum of 3 hours per week of action games exclusively, or if they did play more than 3 hours of video games, but where this total may have included games that did not belong to the action genre, such as when the games listed in the manuscript included role playing, strategy, sports, or fighting games (Adams, 2013, Experiment 2;
Bialystok, 2006;Granek, Gorbet, & Sergio, 2010;Vallett, Lamb, &
Annetta, 2013). In addition, studies in which the NVGPs played more than 1 hour of action games were also excluded (e.g., NVGPs played less than 2 hours of action games and less than 5 hours overall in Dobrowolski, Hanusz, Sobczyk, Skorko, & Wiatrow, 2015; NVGPs played less than 4 hours inDurlach, Kring, & Bowens, 2009). For Rupp, McConnell, and Smither (2016), only the groups of high-play gamers and nonvideo game players met our criteria for inclusion. The group of medium-play gamers was excluded because they played an average of 3 hours of action video games per week suggesting some individuals in this group may have played for fewer than 3 hours.
In another recent study (Unsworth et al., 2015), the authors correlated performance in a number of tasks with hours of action video game play, such that their sample included some AVGPs and NVGPs, but the majority of the individuals were intermediate players. Their data could therefore not be included as published given that our inclusion criteria focused on extreme groups and required some separation between the AVGP and NVGP group in terms of weekly gaming hours. However, thanks to the authors sharing their raw data with us, we were able to compute the effect sizes for the comparison of AVGPs and NVGPs that met our inclusion criteria.
Studies involving clinical populations were systematically ex- cluded (R. W. Li et al., 2011), as were studies comparing action video game players to particular populations such as musicians or bilinguals (Bergstrom, Howard, & Howard, 2012; Bialystok, 2006). Another study was excluded because we could not obtain the data in format corresponding to our action video game players selection criteria (Collins & Freeman, 2014).
3. Game genre for experimental group in intervention studies. For studies involving a training intervention, the exper- imental game had to be an action game as defined above, that is a first or third person shooter game. There was no restriction regard- ing the type of platform (i.e., console, computer/laptop, or mobile device).
4. Game genre for control group in intervention studies.
Training with an action game had to be compared with training on another, nonaction video game. Studies comparing performance before and after training with no control group or with just a ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
test-retest control group were excluded. As is standard in the field, all intervention studies entailed comparisons between a commer- cially available action and a commercially available nonaction video game. This is typical of studies in that domain as it ensures both experimental and control interventions used media known to be engaging and challenging to their users (Jacoby & Ahissar, 2013). Studies in which control training consisted exclusively in repeated practice of a version of the cognitive task used to measure performance changes, such as the UFOV trained group inBelchior et al. (2013), were thus excluded. In addition, two studies com- pared an action game to more than one nonaction video game (Boot, Kramer, Simons, Fabiani, & Gratton, 2008;Oei & Patter- son, 2013), in which case all the comparisons between the action and the variety of nonaction games were included. Finally, only studies in which the experimental and control training had matched schedules were included; for example, Boot, Champion, et al.
(2013)was excluded because the amount of training was highly disparate across training groups.
5. Hours of practice criteria and schedule of testing for intervention studies. Only studies that included a pre- and a posttest were included. Given the focus of the field on long lasting impact of action video game on cognition, a minimum delay of 24 hours was required between the last video game training session and the first posttraining measure of performance. Such a criterion is typical of most studies and rules out the myriad effects that could arise from having just played a video game (e.g., as related to physiological arousal). Studies that used either short training or immediate posttests were thus excluded (e.g.,Gallagher & Preswitch, 2012;Nelson &
Strachan, 2009;Obana & Kozhevnikov, 2012;Rehbein, Kleimann, &
Mossle, 2007; Sanchez, 2012), as well as studies that used only posttest measures (Granek et al., 2010).
Durably changing cognitive skills requires repeated, distributed training; video games are no exception to this (Stafford & Dewar, 2014). Although most studies have used 10 hours or more of training distributed over two weeks, this criterion was relaxed to a minimal duration of 8 hours of training, distributed over a period of at least 8 days.
These combined criteria (delay of 24 hours for posttest and 8 hours of training distributed over 8 days) excluded a few studies:
the ones using training on the scale of just a few minutes (Rehbein et al., 2007) to a few hours (e.g., 4 hours inCherney, 2008; 4 hours on 4 different games inSersale, 2005) as well as studies where training was highly massed across time (e.g.,Gallagher & Pre- switch, 2012did 12 hours in 2 days; van Ravenzwaaij, Boekel, Forstmann, Ratcliff, & Wagenmakers, 2014did 10 hours in 5 days in Experiment 1, and 20 hours in 5 days in Experiment 2), to avoid confounding effects related to the known inefficiency of massed practice (Baddeley & Longman, 1978;Benjamin & Tullis, 2010;
Stafford & Dewar, 2014;Vlach & Sandhofer, 2012).
Among the 82 records included, 32 (39%) included more than one study. Among these, nine records included exactly identical participants, 22 involved independent or partially independent groups of participants, and one included both. Forty-three (37%) of these studies measured performance in more than one domain of cognition. For our analyses, we used a hierarchical model in which the effect sizes are combined at the level of the 116 studies with independent or partially independent subjects over a total of 309 effects. Clusters are therefore defined as effect sizes measures being nested in studies, which allows both within-study and
between-study sources of variances to be considered in the models.
The hierarchical structure of effect sizes and moderators used in our analysis is described in further detail in the section on effect size computation.
Coding of Moderators and Other Study Descriptors For each analysis (cross-sectional and intervention), we provide information relevant to the overall effect of action games on behavior collapsed across cognitive domains and all other moder- ators, as well as a moderator analysis that takes advantage of the methodological heterogeneity, or between-study variability, to provide a more detailed and finer-grained view of the factors modulating the main effect of action video game. To increase the consistency in how the different studies were coded, an initial coding of moderators and study descriptors was performed by Benoit Bediou, who thus had a more complete picture of the data in terms of quality and heterogeneity. Once all studies were coded, the entire dataset—that is all effect sizes, and their classification into cognitive domains, types of studies, type of effect, age group, and so forth—was cross-checked by a minimum of two other authors. We list below for each effect size the moderators coded and included in the analysis (seeTable 1for the number of effect sizes for each moderator level in cross-sectional studies, andTable 2for intervention studies).
Cognitive Domain
We categorized all the computed effect sizes into eight cognitive domains, according to the task and conditions used to extract the relevant effect: (1) perception, (2) bottom-up attention, (3) top- down attention, (4) spatial cognition, (5) multitasking, (6) inhibi- tion, (7) problem solving, and (8) verbal cognition. The number of effect sizes in each cognitive domain for correlational and inter- vention studies can be found respectively inTable 1andTable 2.
The initial selection of relevant task conditions and their clas- sification into cognitive domains described earlier, as well as further coding of the additional moderators considered in the analysis, was first performed by three authors, and then discussed with the rest of the authors until unanimous agreement was reached. For example, the full version of the UFOV task generally contains several conditions that are run in separate blocks, and that probe different skills. The single-task conditions (center task or peripheral task alone) were assigned to the ‘perceptual’ skill; the dual-task condition in the absence of distractors was assigned to the ‘task-switching/multitasking’ skill; and the dual-task condition in the presence of distractors was assigned to the ‘top-down attention’ skill. We recognize that there is a certain degree of arbitrariness in such classifications as clearly the latter condition also qualifies for the ‘multitasking’ skill; yet this task condition is typically utilized for the load it puts on top-down attention. In our classification, we aimed at preserving the intent of the measure in the original work. Each effect size was assigned to one and only one cognitive domain based on the main domain that task com- ponent is hypothesized to tap. The agreement between the three raters involved in the coding of this moderator was good (Cohen’s
⫽.68; Fleiss ⫽.70 and Krippendorf␣ ⫽.70;Cohen, 1960;
Davies & Fleiss, 1982;Fleiss, 1971;Krippendorf, 1980).
Age group. We defined three discrete age ranges, capturing the most important differences in cognitive functioning across the ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
life span: children (below 18, k ⫽ 5 all from cross-sectional studies), younger adults (18 to 35 years old,k⫽ 189 for cross- sectional and 90 for intervention studies) and older adults (above 65,k⫽11 all intervention studies). Three studies (Dye & Bavelier, 2010; Dye, Green, & Bavelier, 2009; Trick, Jaspers-Fayer, &
Sethi, 2005) included groups of children of different ages, with the older group including up to 19-year-olds. Despite including a small number of young adults, these studies were coded as “children”
given that the vast majority of the participants were under 18 years of age.
Type of dependent measure. For each measure, the effect size was classified as reflecting either the speed of processing (e.g., raw or normed RTs, critical stimulus duration, search rates), or its accuracy (e.g., percent correct, error rate, d-prime, threshold). If an effect size was available for both RT and accuracy, only the effect size with the largest absolute magnitude was kept, and classified accordingly.
Main versus difference. Each measure was coded as reflect- ing either a main effect (e.g., overall difference in accuracy, speed, threshold ord-prime), or a difference score (e.g., Flanker compat- ibility effect, difference between single and multitasking condi- tions).
Laboratory of origin. Effect sizes were coded as coming from the group of Bavelier and colleagues, or from other labora- tories.
Recruitment in cross-sectional studies. For each comparison between AVGPs and NVGPs, we coded whether the recruitment of
gaming participants was overt (e.g., via explicit posters or where a video game questionnaire was completed prior to the study), or covert (e.g., a video game questionnaire filled out after the study;
by a third-party such as parents for children; or well prior to the study, as part of a larger battery of questionnaires and where the participant was unaware that the gaming questionnaire was rele- vant to the study at hand). Studies that mixed covert and overt recruitment (e.g.,K. Clark, Fleck, & Mitroff, 2011;Trick et al., 2005) or that did not provide sufficient information to evaluate the type of recruitment were coded as overt.
Training duration in intervention studies. For each inter- vention study, we coded the total duration of training in hours. The median duration was 23.25 hours (M⫽26.20, range 10 –50 hours).
Effect Size Computation
Effect sizes give the magnitude and the direction either of the difference between two groups (i.e., AVGPs vs. NVGPs) on a given measure, or of the difference in the magnitude of pretest to posttest changes between two treatments (i.e., training on an action vs. on a nonaction control game). Positive effect sizes reflect greater performance (or improvement) in AVGPs (or action- trained) compared with NVGPs (or control-trained) groups (i.e., more correct responses, fewer errors, or faster RTs).
We used the bias-corrected Hedges’ g as our main measure of effect sizes. This is equivalent to a Cohen’sdwith an additional correction factor for small samples, and is thus more conservative Table 1
Cross-Sectional Meta-Analysis
Moderator–Level k m F g 95% CI df p
Cognitive domain 194 89 2.125 7.9 .161
Perception 30 22 .775 [.564, .985] 16.6 ⬍.001ⴱⴱⴱ
Top-down attention 71 48 .625 [.494, .756] 27.4 ⬍.001ⴱⴱⴱ
Spatial cognition 27 19 .750 [.526, .975] 14.5 ⬍.001ⴱⴱⴱ
Inhibition 11 9 .310 [.065, .556] 7.2 .02ⴱ
Multi-tasking 22 17 .549 [.277, .821] 11.9 ⬍.001ⴱⴱⴱ
Problem solving 7 4 .501 [⫺.017, 1.019] 2.4 .054ⴱ
Verbal cognition 26 16 .297 [.032, .563] 7.7 .033ⴱ
Age group 194 89 1.652 3.2 .283
Children 5 3 .324 [⫺.337, .985] 2.9 .21
Younger adults 189 86 .598 [.498, .697] 33.6 ⬍.001ⴱⴱⴱ
DV type 194 89 .140 35.4 .711
Accuracy 139 61 .582 [.467, .697] 31.2 ⬍.001ⴱⴱⴱ
Speed 55 41 .614 [.467, .76] 31.5 ⬍.001ⴱⴱⴱ
Effect type 194 89 1.560 23.1 .224
Main 139 65 .620 [.515, .724] 33.4 ⬍.001ⴱⴱⴱ
Difference 55 35 .518 [.353, .683] 17.3 ⬍.001ⴱⴱⴱ
Lab 194 89 4.657 26.8 .040ⴱ
Bavelier 54 32 .800 [.551, 1.05] 19.5 ⬍.001ⴱⴱⴱ
Other 140 57 .510 [.404, .616] 29.1 ⬍.001ⴱⴱⴱ
Recruitment 194 89 1.616 13.2 .226
Overt 140 74 .624 [.504, .745] 33 ⬍.001ⴱⴱⴱ
Covert 54 16 .504 [.329, .679] 7.2 ⬍.001ⴱⴱⴱ
Note. The effect sizes of each moderator level are shown; these analyses are based on robust variance estimates with a model including all moderators.k⫽number of effect sizes;m⫽number of clusters (Ftests) or individual studies (ttests) for each moderator and each level, respectively.F⫽AHT-Ftest comparing the levels of a given moderator.g⫽effect size estimate (Hedges); 95% CI⫽95% confidence interval;df⫽degrees of freedom;p⫽ pvalue of AHT-Ftests for moderator effects andttests comparing each level against zero. For each moderator, the first row (in bold and italics) shows the result of theFtests (AHT Type) examining possible differences between the levels of each moderator.
ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.
